Slices, for efficient multi-tenanancy

In environments in which data is naturally segregated (such as multi-tenant setups), creating an individual index per tenant creates administrative and computing overhead. Conversely, filtering a shared index by tenant ID is an additional effort during search and could impact aspects like query latency and recall.

Slices provides a simple way to organize the data in the index by the tenant ID (or by any other field that defines how the data will typically be consumed). Elasticsearch will route data from a slice to a specific shard and sort the documents by slice in the index segment. This ensures queries efficiently scan only the relevant parts of the index. There are additional optimizations using slice for vector search when using DiskBBQ.

This results in faster queries, higher ranking quality, and reduced compute costs by potentially requiring fewer replicas.